wllama-service 1.0.0

package/README.md

@@ -0,0 +1,56 @@
+# wllama-service
+
+A framework-agnostic browser LLM service wrapper built on [@wllama/wllama](https://github.com/ngxson/wllama).
+Runs GGUF models locally in the browser with WebGPU acceleration and WebAssembly fallback.
+
+## Requirements
+
+Your web server must send these headers for multi-threading and WebGPU to work:
+```
+Cross-Origin-Opener-Policy: same-origin
+Cross-Origin-Embedder-Policy: require-corp
+```
+
+Copy `wllama.wasm` from `node_modules/@wllama/wllama/esm/wasm/wllama.wasm` to your public folder
+and pass its URL via `wasmPath`.
+
+## Install
+
+```bash
+npm install wllama-service @wllama/wllama
+```
+
+## Usage
+
+```typescript
+import { WllamaService } from 'wllama-service';
+
+const service = new WllamaService({
+  wasmPath: '/wllama/wllama.wasm', // path in your public folder
+  nCtx: 2048,
+  nGpuLayers: 999, // set 0 to disable WebGPU
+});
+
+// Check browser support
+const env = service.checkEnvironment();
+console.log('WebGPU available:', env.hasWebGPU);
+
+// Load a GGUF file from file input
+const result = await service.loadModel(file, (progress) => {
+  console.log(`Loading: ${progress}%`);
+});
+
+// Generate
+const response = await service.generate({
+  system: 'You are a helpful assistant.',
+  prompt: 'Hello!',
+  maxTokens: 256,
+  temperature: 0.7,
+});
+
+console.log(response.text);
+console.log(`Generated in ${response.timeMs}ms`);
+
+// Unload when done
+await service.unload();
+```

package/dist/index.d.mts

@@ -0,0 +1,58 @@
+interface WllamaServiceConfig {
+    /** Public URL where wllama.wasm is served from. Default: '/wllama/wllama.wasm' */
+    wasmPath?: string;
+    /** Number of GPU layers to offload. Default: 999 (all). Set 0 to disable WebGPU. */
+    nGpuLayers?: number;
+    /** Context window size. Default: 2048 */
+    nCtx?: number;
+}
+interface GenerateRequest {
+    prompt: string;
+    system?: string;
+    maxTokens?: number;
+    temperature?: number;
+    topK?: number;
+    topP?: number;
+}
+interface GenerateResult {
+    success: boolean;
+    text?: string;
+    error?: string;
+    timeMs?: number;
+}
+interface LoadModelResult {
+    success: boolean;
+    error?: string;
+    usedWebGPU?: boolean;
+}
+interface ConnectionResult {
+    success: boolean;
+    error?: string;
+    hasWebGPU?: boolean;
+    crossOriginIsolated?: boolean;
+}
+
+declare class WllamaService {
+    private wllama;
+    private modelName;
+    private config;
+    constructor(config?: WllamaServiceConfig);
+    /** Check browser compatibility and capabilities */
+    checkEnvironment(): ConnectionResult;
+    /** Load a GGUF model file from the user's filesystem */
+    loadModel(file: File, onProgress?: (progress: number) => void): Promise<LoadModelResult>;
+    /** Generate a response from the loaded model */
+    generate(req: GenerateRequest): Promise<GenerateResult>;
+    /** Unload the current model and free memory */
+    unload(): Promise<void>;
+    /** Whether a model is currently loaded */
+    get isLoaded(): boolean;
+    /** Name of the currently loaded model */
+    get currentModel(): string;
+    private storeInCache;
+    private deleteFromCache;
+}
+
+declare function copyWasmToPublic(publicDir: string): void;
+
+export { type ConnectionResult, type GenerateRequest, type GenerateResult, type LoadModelResult, WllamaService, type WllamaServiceConfig, copyWasmToPublic };

package/dist/index.d.ts

@@ -0,0 +1,58 @@
+interface WllamaServiceConfig {
+    /** Public URL where wllama.wasm is served from. Default: '/wllama/wllama.wasm' */
+    wasmPath?: string;
+    /** Number of GPU layers to offload. Default: 999 (all). Set 0 to disable WebGPU. */
+    nGpuLayers?: number;
+    /** Context window size. Default: 2048 */
+    nCtx?: number;
+}
+interface GenerateRequest {
+    prompt: string;
+    system?: string;
+    maxTokens?: number;
+    temperature?: number;
+    topK?: number;
+    topP?: number;
+}
+interface GenerateResult {
+    success: boolean;
+    text?: string;
+    error?: string;
+    timeMs?: number;
+}
+interface LoadModelResult {
+    success: boolean;
+    error?: string;
+    usedWebGPU?: boolean;
+}
+interface ConnectionResult {
+    success: boolean;
+    error?: string;
+    hasWebGPU?: boolean;
+    crossOriginIsolated?: boolean;
+}
+
+declare class WllamaService {
+    private wllama;
+    private modelName;
+    private config;
+    constructor(config?: WllamaServiceConfig);
+    /** Check browser compatibility and capabilities */
+    checkEnvironment(): ConnectionResult;
+    /** Load a GGUF model file from the user's filesystem */
+    loadModel(file: File, onProgress?: (progress: number) => void): Promise<LoadModelResult>;
+    /** Generate a response from the loaded model */
+    generate(req: GenerateRequest): Promise<GenerateResult>;
+    /** Unload the current model and free memory */
+    unload(): Promise<void>;
+    /** Whether a model is currently loaded */
+    get isLoaded(): boolean;
+    /** Name of the currently loaded model */
+    get currentModel(): string;
+    private storeInCache;
+    private deleteFromCache;
+}
+
+declare function copyWasmToPublic(publicDir: string): void;
+
+export { type ConnectionResult, type GenerateRequest, type GenerateResult, type LoadModelResult, WllamaService, type WllamaServiceConfig, copyWasmToPublic };

package/dist/index.js

@@ -0,0 +1,175 @@
+"use strict";
+var __create = Object.create;
+var __defProp = Object.defineProperty;
+var __getOwnPropDesc = Object.getOwnPropertyDescriptor;
+var __getOwnPropNames = Object.getOwnPropertyNames;
+var __getProtoOf = Object.getPrototypeOf;
+var __hasOwnProp = Object.prototype.hasOwnProperty;
+var __export = (target, all) => {
+  for (var name in all)
+    __defProp(target, name, { get: all[name], enumerable: true });
+};
+var __copyProps = (to, from, except, desc) => {
+  if (from && typeof from === "object" || typeof from === "function") {
+    for (let key of __getOwnPropNames(from))
+      if (!__hasOwnProp.call(to, key) && key !== except)
+        __defProp(to, key, { get: () => from[key], enumerable: !(desc = __getOwnPropDesc(from, key)) || desc.enumerable });
+  }
+  return to;
+};
+var __toESM = (mod, isNodeMode, target) => (target = mod != null ? __create(__getProtoOf(mod)) : {}, __copyProps(
+  // If the importer is in node compatibility mode or this is not an ESM
+  // file that has been converted to a CommonJS file using a Babel-
+  // compatible transform (i.e. "__esModule" has not been set), then set
+  // "default" to the CommonJS "module.exports" for node compatibility.
+  isNodeMode || !mod || !mod.__esModule ? __defProp(target, "default", { value: mod, enumerable: true }) : target,
+  mod
+));
+var __toCommonJS = (mod) => __copyProps(__defProp({}, "__esModule", { value: true }), mod);
+
+// src/index.ts
+var index_exports = {};
+__export(index_exports, {
+  WllamaService: () => WllamaService,
+  copyWasmToPublic: () => copyWasmToPublic
+});
+module.exports = __toCommonJS(index_exports);
+
+// src/WllamaService.ts
+var import_esm = require("@wllama/wllama/esm/index.js");
+var WllamaService = class {
+  constructor(config = {}) {
+    this.wllama = null;
+    this.modelName = "";
+    this.config = {
+      wasmPath: config.wasmPath ?? "/wllama/wllama.wasm",
+      nGpuLayers: config.nGpuLayers ?? 999,
+      nCtx: config.nCtx ?? 2048
+    };
+  }
+  /** Check browser compatibility and capabilities */
+  checkEnvironment() {
+    if (typeof WebAssembly === "undefined") {
+      return { success: false, error: "WebAssembly is not supported in this browser." };
+    }
+    if (!("caches" in window)) {
+      return { success: false, error: "Cache API is not supported in this browser." };
+    }
+    const hasWebGPU = !!navigator.gpu;
+    const isolated = window.crossOriginIsolated;
+    return {
+      success: true,
+      hasWebGPU,
+      crossOriginIsolated: isolated
+    };
+  }
+  /** Load a GGUF model file from the user's filesystem */
+  async loadModel(file, onProgress) {
+    const env = this.checkEnvironment();
+    if (!env.success) return { success: false, error: env.error };
+    await this.unload();
+    try {
+      onProgress?.(10);
+      const configPaths = { default: this.config.wasmPath };
+      this.wllama = new import_esm.Wllama(configPaths);
+      onProgress?.(30);
+      await this.storeInCache(file);
+      onProgress?.(50);
+      const hasWebGPU = !!navigator.gpu;
+      await this.wllama.loadModel([file], {
+        n_ctx: this.config.nCtx,
+        ...hasWebGPU && this.config.nGpuLayers > 0 ? { n_gpu_layers: this.config.nGpuLayers } : {}
+      });
+      this.modelName = file.name;
+      onProgress?.(100);
+      return { success: true, usedWebGPU: hasWebGPU };
+    } catch (e) {
+      await this.unload();
+      return { success: false, error: e.message };
+    }
+  }
+  /** Generate a response from the loaded model */
+  async generate(req) {
+    if (!this.wllama) {
+      return { success: false, error: "No model loaded. Call loadModel() first." };
+    }
+    try {
+      const t0 = performance.now();
+      const messages = [];
+      if (req.system) messages.push({ role: "system", content: req.system });
+      messages.push({ role: "user", content: req.prompt });
+      const response = await this.wllama.createChatCompletion({
+        messages,
+        max_tokens: req.maxTokens ?? 512,
+        temperature: req.temperature ?? 0.7,
+        top_k: req.topK ?? 40,
+        top_p: req.topP ?? 0.95
+      });
+      const text = response?.choices?.[0]?.message?.content ?? "";
+      return {
+        success: true,
+        text,
+        timeMs: Math.round(performance.now() - t0)
+      };
+    } catch (e) {
+      return { success: false, error: e.message || "Generation failed" };
+    }
+  }
+  /** Unload the current model and free memory */
+  async unload() {
+    if (this.wllama) {
+      try {
+        await this.wllama.exit();
+      } catch (_) {
+      }
+      this.wllama = null;
+    }
+    if (this.modelName) {
+      try {
+        await this.deleteFromCache(this.modelName);
+      } catch (_) {
+      }
+      this.modelName = "";
+    }
+  }
+  /** Whether a model is currently loaded */
+  get isLoaded() {
+    return this.wllama !== null;
+  }
+  /** Name of the currently loaded model */
+  get currentModel() {
+    return this.modelName;
+  }
+  async storeInCache(file) {
+    const url = `${window.location.origin}/wllama-local-models/${file.name}`;
+    const cache = await caches.open("wllama-local-models");
+    await cache.put(url, new Response(file, {
+      headers: { "Content-Type": "application/octet-stream" }
+    }));
+  }
+  async deleteFromCache(fileName) {
+    const url = `${window.location.origin}/wllama-local-models/${fileName}`;
+    const cache = await caches.open("wllama-local-models");
+    await cache.delete(url);
+  }
+};
+
+// src/copy-wasm.ts
+var import_fs = __toESM(require("fs"));
+var import_path = __toESM(require("path"));
+function copyWasmToPublic(publicDir) {
+  const src = import_path.default.resolve(
+    require.resolve("@wllama/wllama/esm/index.js"),
+    "../../wasm/wllama.wasm"
+  );
+  const dest = import_path.default.join(publicDir, "wllama", "wllama.wasm");
+  import_fs.default.mkdirSync(import_path.default.dirname(dest), { recursive: true });
+  import_fs.default.copyFileSync(src, dest);
+  console.log(`\u2713 Copied wllama.wasm to ${dest}`);
+}
+// Annotate the CommonJS export names for ESM import in node:
+0 && (module.exports = {
+  WllamaService,
+  copyWasmToPublic
+});
+//# sourceMappingURL=index.js.map

package/dist/index.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/index.ts","../src/WllamaService.ts","../src/copy-wasm.ts"],"sourcesContent":["export { WllamaService } from './WllamaService';\nexport type {\n  WllamaServiceConfig,\n  GenerateRequest,\n  GenerateResult,\n  LoadModelResult,\n  ConnectionResult,\n} from './types';\nexport { copyWasmToPublic } from './copy-wasm';","import { Wllama } from '@wllama/wllama/esm/index.js';\nimport type {\n  WllamaServiceConfig,\n  GenerateRequest,\n  GenerateResult,\n  LoadModelResult,\n  ConnectionResult,\n} from './types';\n\nexport class WllamaService {\n  private wllama: Wllama | null = null;\n  private modelName: string = '';\n  private config: Required<WllamaServiceConfig>;\n\n  constructor(config: WllamaServiceConfig = {}) {\n    this.config = {\n      wasmPath: config.wasmPath ?? '/wllama/wllama.wasm',\n      nGpuLayers: config.nGpuLayers ?? 999,\n      nCtx: config.nCtx ?? 2048,\n    };\n  }\n\n  /** Check browser compatibility and capabilities */\n  checkEnvironment(): ConnectionResult {\n    if (typeof WebAssembly === 'undefined') {\n      return { success: false, error: 'WebAssembly is not supported in this browser.' };\n    }\n    if (!('caches' in window)) {\n      return { success: false, error: 'Cache API is not supported in this browser.' };\n    }\n    const hasWebGPU = !!(navigator as any).gpu;\n    const isolated = window.crossOriginIsolated;\n    return {\n      success: true,\n      hasWebGPU,\n      crossOriginIsolated: isolated,\n    };\n  }\n\n  /** Load a GGUF model file from the user's filesystem */\n  async loadModel(\n    file: File,\n    onProgress?: (progress: number) => void,\n  ): Promise<LoadModelResult> {\n    const env = this.checkEnvironment();\n    if (!env.success) return { success: false, error: env.error };\n\n    await this.unload();\n\n    try {\n      onProgress?.(10);\n\n      const configPaths = { default: this.config.wasmPath };\n      this.wllama = new Wllama(configPaths as any);\n      onProgress?.(30);\n\n      await this.storeInCache(file);\n      onProgress?.(50);\n\n      const hasWebGPU = !!(navigator as any).gpu;\n      await (this.wllama as any).loadModel([file], {\n        n_ctx: this.config.nCtx,\n        ...(hasWebGPU && this.config.nGpuLayers > 0\n          ? { n_gpu_layers: this.config.nGpuLayers }\n          : {}),\n      });\n\n      this.modelName = file.name;\n      onProgress?.(100);\n\n      return { success: true, usedWebGPU: hasWebGPU };\n    } catch (e: any) {\n      await this.unload();\n      return { success: false, error: e.message };\n    }\n  }\n\n  /** Generate a response from the loaded model */\n  async generate(req: GenerateRequest): Promise<GenerateResult> {\n    if (!this.wllama) {\n      return { success: false, error: 'No model loaded. Call loadModel() first.' };\n    }\n\n    try {\n      const t0 = performance.now();\n\n      const messages: { role: string; content: string }[] = [];\n      if (req.system) messages.push({ role: 'system', content: req.system });\n      messages.push({ role: 'user', content: req.prompt });\n\n      const response = await (this.wllama as any).createChatCompletion({\n        messages,\n        max_tokens: req.maxTokens ?? 512,\n        temperature: req.temperature ?? 0.7,\n        top_k: req.topK ?? 40,\n        top_p: req.topP ?? 0.95,\n      });\n\n      const text: string = response?.choices?.[0]?.message?.content ?? '';\n      return {\n        success: true,\n        text,\n        timeMs: Math.round(performance.now() - t0),\n      };\n    } catch (e: any) {\n      return { success: false, error: e.message || 'Generation failed' };\n    }\n  }\n\n  /** Unload the current model and free memory */\n  async unload(): Promise<void> {\n    if (this.wllama) {\n      try { await (this.wllama as any).exit(); } catch (_) {}\n      this.wllama = null;\n    }\n    if (this.modelName) {\n      try { await this.deleteFromCache(this.modelName); } catch (_) {}\n      this.modelName = '';\n    }\n  }\n\n  /** Whether a model is currently loaded */\n  get isLoaded(): boolean {\n    return this.wllama !== null;\n  }\n\n  /** Name of the currently loaded model */\n  get currentModel(): string {\n    return this.modelName;\n  }\n\n  private async storeInCache(file: File): Promise<void> {\n    const url = `${window.location.origin}/wllama-local-models/${file.name}`;\n    const cache = await caches.open('wllama-local-models');\n    await cache.put(url, new Response(file, {\n      headers: { 'Content-Type': 'application/octet-stream' },\n    }));\n  }\n\n  private async deleteFromCache(fileName: string): Promise<void> {\n    const url = `${window.location.origin}/wllama-local-models/${fileName}`;\n    const cache = await caches.open('wllama-local-models');\n    await cache.delete(url);\n  }\n}","import fs from 'fs';\nimport path from 'path';\n\nexport function copyWasmToPublic(publicDir: string): void {\n  const src = path.resolve(\n    require.resolve('@wllama/wllama/esm/index.js'),\n    '../../wasm/wllama.wasm'\n  );\n  const dest = path.join(publicDir, 'wllama', 'wllama.wasm');\n  \n  fs.mkdirSync(path.dirname(dest), { recursive: true });\n  fs.copyFileSync(src, dest);\n  console.log(`✓ Copied wllama.wasm to ${dest}`);\n}"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;;;ACAA,iBAAuB;AAShB,IAAM,gBAAN,MAAoB;AAAA,EAKzB,YAAY,SAA8B,CAAC,GAAG;AAJ9C,SAAQ,SAAwB;AAChC,SAAQ,YAAoB;AAI1B,SAAK,SAAS;AAAA,MACZ,UAAU,OAAO,YAAY;AAAA,MAC7B,YAAY,OAAO,cAAc;AAAA,MACjC,MAAM,OAAO,QAAQ;AAAA,IACvB;AAAA,EACF;AAAA;AAAA,EAGA,mBAAqC;AACnC,QAAI,OAAO,gBAAgB,aAAa;AACtC,aAAO,EAAE,SAAS,OAAO,OAAO,gDAAgD;AAAA,IAClF;AACA,QAAI,EAAE,YAAY,SAAS;AACzB,aAAO,EAAE,SAAS,OAAO,OAAO,8CAA8C;AAAA,IAChF;AACA,UAAM,YAAY,CAAC,CAAE,UAAkB;AACvC,UAAM,WAAW,OAAO;AACxB,WAAO;AAAA,MACL,SAAS;AAAA,MACT;AAAA,MACA,qBAAqB;AAAA,IACvB;AAAA,EACF;AAAA;AAAA,EAGA,MAAM,UACJ,MACA,YAC0B;AAC1B,UAAM,MAAM,KAAK,iBAAiB;AAClC,QAAI,CAAC,IAAI,QAAS,QAAO,EAAE,SAAS,OAAO,OAAO,IAAI,MAAM;AAE5D,UAAM,KAAK,OAAO;AAElB,QAAI;AACF,mBAAa,EAAE;AAEf,YAAM,cAAc,EAAE,SAAS,KAAK,OAAO,SAAS;AACpD,WAAK,SAAS,IAAI,kBAAO,WAAkB;AAC3C,mBAAa,EAAE;AAEf,YAAM,KAAK,aAAa,IAAI;AAC5B,mBAAa,EAAE;AAEf,YAAM,YAAY,CAAC,CAAE,UAAkB;AACvC,YAAO,KAAK,OAAe,UAAU,CAAC,IAAI,GAAG;AAAA,QAC3C,OAAO,KAAK,OAAO;AAAA,QACnB,GAAI,aAAa,KAAK,OAAO,aAAa,IACtC,EAAE,cAAc,KAAK,OAAO,WAAW,IACvC,CAAC;AAAA,MACP,CAAC;AAED,WAAK,YAAY,KAAK;AACtB,mBAAa,GAAG;AAEhB,aAAO,EAAE,SAAS,MAAM,YAAY,UAAU;AAAA,IAChD,SAAS,GAAQ;AACf,YAAM,KAAK,OAAO;AAClB,aAAO,EAAE,SAAS,OAAO,OAAO,EAAE,QAAQ;AAAA,IAC5C;AAAA,EACF;AAAA;AAAA,EAGA,MAAM,SAAS,KAA+C;AAC5D,QAAI,CAAC,KAAK,QAAQ;AAChB,aAAO,EAAE,SAAS,OAAO,OAAO,2CAA2C;AAAA,IAC7E;AAEA,QAAI;AACF,YAAM,KAAK,YAAY,IAAI;AAE3B,YAAM,WAAgD,CAAC;AACvD,UAAI,IAAI,OAAQ,UAAS,KAAK,EAAE,MAAM,UAAU,SAAS,IAAI,OAAO,CAAC;AACrE,eAAS,KAAK,EAAE,MAAM,QAAQ,SAAS,IAAI,OAAO,CAAC;AAEnD,YAAM,WAAW,MAAO,KAAK,OAAe,qBAAqB;AAAA,QAC/D;AAAA,QACA,YAAY,IAAI,aAAa;AAAA,QAC7B,aAAa,IAAI,eAAe;AAAA,QAChC,OAAO,IAAI,QAAQ;AAAA,QACnB,OAAO,IAAI,QAAQ;AAAA,MACrB,CAAC;AAED,YAAM,OAAe,UAAU,UAAU,CAAC,GAAG,SAAS,WAAW;AACjE,aAAO;AAAA,QACL,SAAS;AAAA,QACT;AAAA,QACA,QAAQ,KAAK,MAAM,YAAY,IAAI,IAAI,EAAE;AAAA,MAC3C;AAAA,IACF,SAAS,GAAQ;AACf,aAAO,EAAE,SAAS,OAAO,OAAO,EAAE,WAAW,oBAAoB;AAAA,IACnE;AAAA,EACF;AAAA;AAAA,EAGA,MAAM,SAAwB;AAC5B,QAAI,KAAK,QAAQ;AACf,UAAI;AAAE,cAAO,KAAK,OAAe,KAAK;AAAA,MAAG,SAAS,GAAG;AAAA,MAAC;AACtD,WAAK,SAAS;AAAA,IAChB;AACA,QAAI,KAAK,WAAW;AAClB,UAAI;AAAE,cAAM,KAAK,gBAAgB,KAAK,SAAS;AAAA,MAAG,SAAS,GAAG;AAAA,MAAC;AAC/D,WAAK,YAAY;AAAA,IACnB;AAAA,EACF;AAAA;AAAA,EAGA,IAAI,WAAoB;AACtB,WAAO,KAAK,WAAW;AAAA,EACzB;AAAA;AAAA,EAGA,IAAI,eAAuB;AACzB,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,MAAc,aAAa,MAA2B;AACpD,UAAM,MAAM,GAAG,OAAO,SAAS,MAAM,wBAAwB,KAAK,IAAI;AACtE,UAAM,QAAQ,MAAM,OAAO,KAAK,qBAAqB;AACrD,UAAM,MAAM,IAAI,KAAK,IAAI,SAAS,MAAM;AAAA,MACtC,SAAS,EAAE,gBAAgB,2BAA2B;AAAA,IACxD,CAAC,CAAC;AAAA,EACJ;AAAA,EAEA,MAAc,gBAAgB,UAAiC;AAC7D,UAAM,MAAM,GAAG,OAAO,SAAS,MAAM,wBAAwB,QAAQ;AACrE,UAAM,QAAQ,MAAM,OAAO,KAAK,qBAAqB;AACrD,UAAM,MAAM,OAAO,GAAG;AAAA,EACxB;AACF;;;AChJA,gBAAe;AACf,kBAAiB;AAEV,SAAS,iBAAiB,WAAyB;AACxD,QAAM,MAAM,YAAAA,QAAK;AAAA,IACf,gBAAgB,6BAA6B;AAAA,IAC7C;AAAA,EACF;AACA,QAAM,OAAO,YAAAA,QAAK,KAAK,WAAW,UAAU,aAAa;AAEzD,YAAAC,QAAG,UAAU,YAAAD,QAAK,QAAQ,IAAI,GAAG,EAAE,WAAW,KAAK,CAAC;AACpD,YAAAC,QAAG,aAAa,KAAK,IAAI;AACzB,UAAQ,IAAI,gCAA2B,IAAI,EAAE;AAC/C;","names":["path","fs"]}

package/dist/index.mjs

@@ -0,0 +1,144 @@
+var __require = /* @__PURE__ */ ((x) => typeof require !== "undefined" ? require : typeof Proxy !== "undefined" ? new Proxy(x, {
+  get: (a, b) => (typeof require !== "undefined" ? require : a)[b]
+}) : x)(function(x) {
+  if (typeof require !== "undefined") return require.apply(this, arguments);
+  throw Error('Dynamic require of "' + x + '" is not supported');
+});
+
+// src/WllamaService.ts
+import { Wllama } from "@wllama/wllama/esm/index.js";
+var WllamaService = class {
+  constructor(config = {}) {
+    this.wllama = null;
+    this.modelName = "";
+    this.config = {
+      wasmPath: config.wasmPath ?? "/wllama/wllama.wasm",
+      nGpuLayers: config.nGpuLayers ?? 999,
+      nCtx: config.nCtx ?? 2048
+    };
+  }
+  /** Check browser compatibility and capabilities */
+  checkEnvironment() {
+    if (typeof WebAssembly === "undefined") {
+      return { success: false, error: "WebAssembly is not supported in this browser." };
+    }
+    if (!("caches" in window)) {
+      return { success: false, error: "Cache API is not supported in this browser." };
+    }
+    const hasWebGPU = !!navigator.gpu;
+    const isolated = window.crossOriginIsolated;
+    return {
+      success: true,
+      hasWebGPU,
+      crossOriginIsolated: isolated
+    };
+  }
+  /** Load a GGUF model file from the user's filesystem */
+  async loadModel(file, onProgress) {
+    const env = this.checkEnvironment();
+    if (!env.success) return { success: false, error: env.error };
+    await this.unload();
+    try {
+      onProgress?.(10);
+      const configPaths = { default: this.config.wasmPath };
+      this.wllama = new Wllama(configPaths);
+      onProgress?.(30);
+      await this.storeInCache(file);
+      onProgress?.(50);
+      const hasWebGPU = !!navigator.gpu;
+      await this.wllama.loadModel([file], {
+        n_ctx: this.config.nCtx,
+        ...hasWebGPU && this.config.nGpuLayers > 0 ? { n_gpu_layers: this.config.nGpuLayers } : {}
+      });
+      this.modelName = file.name;
+      onProgress?.(100);
+      return { success: true, usedWebGPU: hasWebGPU };
+    } catch (e) {
+      await this.unload();
+      return { success: false, error: e.message };
+    }
+  }
+  /** Generate a response from the loaded model */
+  async generate(req) {
+    if (!this.wllama) {
+      return { success: false, error: "No model loaded. Call loadModel() first." };
+    }
+    try {
+      const t0 = performance.now();
+      const messages = [];
+      if (req.system) messages.push({ role: "system", content: req.system });
+      messages.push({ role: "user", content: req.prompt });
+      const response = await this.wllama.createChatCompletion({
+        messages,
+        max_tokens: req.maxTokens ?? 512,
+        temperature: req.temperature ?? 0.7,
+        top_k: req.topK ?? 40,
+        top_p: req.topP ?? 0.95
+      });
+      const text = response?.choices?.[0]?.message?.content ?? "";
+      return {
+        success: true,
+        text,
+        timeMs: Math.round(performance.now() - t0)
+      };
+    } catch (e) {
+      return { success: false, error: e.message || "Generation failed" };
+    }
+  }
+  /** Unload the current model and free memory */
+  async unload() {
+    if (this.wllama) {
+      try {
+        await this.wllama.exit();
+      } catch (_) {
+      }
+      this.wllama = null;
+    }
+    if (this.modelName) {
+      try {
+        await this.deleteFromCache(this.modelName);
+      } catch (_) {
+      }
+      this.modelName = "";
+    }
+  }
+  /** Whether a model is currently loaded */
+  get isLoaded() {
+    return this.wllama !== null;
+  }
+  /** Name of the currently loaded model */
+  get currentModel() {
+    return this.modelName;
+  }
+  async storeInCache(file) {
+    const url = `${window.location.origin}/wllama-local-models/${file.name}`;
+    const cache = await caches.open("wllama-local-models");
+    await cache.put(url, new Response(file, {
+      headers: { "Content-Type": "application/octet-stream" }
+    }));
+  }
+  async deleteFromCache(fileName) {
+    const url = `${window.location.origin}/wllama-local-models/${fileName}`;
+    const cache = await caches.open("wllama-local-models");
+    await cache.delete(url);
+  }
+};
+
+// src/copy-wasm.ts
+import fs from "fs";
+import path from "path";
+function copyWasmToPublic(publicDir) {
+  const src = path.resolve(
+    __require.resolve("@wllama/wllama/esm/index.js"),
+    "../../wasm/wllama.wasm"
+  );
+  const dest = path.join(publicDir, "wllama", "wllama.wasm");
+  fs.mkdirSync(path.dirname(dest), { recursive: true });
+  fs.copyFileSync(src, dest);
+  console.log(`\u2713 Copied wllama.wasm to ${dest}`);
+}
+export {
+  WllamaService,
+  copyWasmToPublic
+};
+//# sourceMappingURL=index.mjs.map

package/dist/index.mjs.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/WllamaService.ts","../src/copy-wasm.ts"],"sourcesContent":["import { Wllama } from '@wllama/wllama/esm/index.js';\nimport type {\n  WllamaServiceConfig,\n  GenerateRequest,\n  GenerateResult,\n  LoadModelResult,\n  ConnectionResult,\n} from './types';\n\nexport class WllamaService {\n  private wllama: Wllama | null = null;\n  private modelName: string = '';\n  private config: Required<WllamaServiceConfig>;\n\n  constructor(config: WllamaServiceConfig = {}) {\n    this.config = {\n      wasmPath: config.wasmPath ?? '/wllama/wllama.wasm',\n      nGpuLayers: config.nGpuLayers ?? 999,\n      nCtx: config.nCtx ?? 2048,\n    };\n  }\n\n  /** Check browser compatibility and capabilities */\n  checkEnvironment(): ConnectionResult {\n    if (typeof WebAssembly === 'undefined') {\n      return { success: false, error: 'WebAssembly is not supported in this browser.' };\n    }\n    if (!('caches' in window)) {\n      return { success: false, error: 'Cache API is not supported in this browser.' };\n    }\n    const hasWebGPU = !!(navigator as any).gpu;\n    const isolated = window.crossOriginIsolated;\n    return {\n      success: true,\n      hasWebGPU,\n      crossOriginIsolated: isolated,\n    };\n  }\n\n  /** Load a GGUF model file from the user's filesystem */\n  async loadModel(\n    file: File,\n    onProgress?: (progress: number) => void,\n  ): Promise<LoadModelResult> {\n    const env = this.checkEnvironment();\n    if (!env.success) return { success: false, error: env.error };\n\n    await this.unload();\n\n    try {\n      onProgress?.(10);\n\n      const configPaths = { default: this.config.wasmPath };\n      this.wllama = new Wllama(configPaths as any);\n      onProgress?.(30);\n\n      await this.storeInCache(file);\n      onProgress?.(50);\n\n      const hasWebGPU = !!(navigator as any).gpu;\n      await (this.wllama as any).loadModel([file], {\n        n_ctx: this.config.nCtx,\n        ...(hasWebGPU && this.config.nGpuLayers > 0\n          ? { n_gpu_layers: this.config.nGpuLayers }\n          : {}),\n      });\n\n      this.modelName = file.name;\n      onProgress?.(100);\n\n      return { success: true, usedWebGPU: hasWebGPU };\n    } catch (e: any) {\n      await this.unload();\n      return { success: false, error: e.message };\n    }\n  }\n\n  /** Generate a response from the loaded model */\n  async generate(req: GenerateRequest): Promise<GenerateResult> {\n    if (!this.wllama) {\n      return { success: false, error: 'No model loaded. Call loadModel() first.' };\n    }\n\n    try {\n      const t0 = performance.now();\n\n      const messages: { role: string; content: string }[] = [];\n      if (req.system) messages.push({ role: 'system', content: req.system });\n      messages.push({ role: 'user', content: req.prompt });\n\n      const response = await (this.wllama as any).createChatCompletion({\n        messages,\n        max_tokens: req.maxTokens ?? 512,\n        temperature: req.temperature ?? 0.7,\n        top_k: req.topK ?? 40,\n        top_p: req.topP ?? 0.95,\n      });\n\n      const text: string = response?.choices?.[0]?.message?.content ?? '';\n      return {\n        success: true,\n        text,\n        timeMs: Math.round(performance.now() - t0),\n      };\n    } catch (e: any) {\n      return { success: false, error: e.message || 'Generation failed' };\n    }\n  }\n\n  /** Unload the current model and free memory */\n  async unload(): Promise<void> {\n    if (this.wllama) {\n      try { await (this.wllama as any).exit(); } catch (_) {}\n      this.wllama = null;\n    }\n    if (this.modelName) {\n      try { await this.deleteFromCache(this.modelName); } catch (_) {}\n      this.modelName = '';\n    }\n  }\n\n  /** Whether a model is currently loaded */\n  get isLoaded(): boolean {\n    return this.wllama !== null;\n  }\n\n  /** Name of the currently loaded model */\n  get currentModel(): string {\n    return this.modelName;\n  }\n\n  private async storeInCache(file: File): Promise<void> {\n    const url = `${window.location.origin}/wllama-local-models/${file.name}`;\n    const cache = await caches.open('wllama-local-models');\n    await cache.put(url, new Response(file, {\n      headers: { 'Content-Type': 'application/octet-stream' },\n    }));\n  }\n\n  private async deleteFromCache(fileName: string): Promise<void> {\n    const url = `${window.location.origin}/wllama-local-models/${fileName}`;\n    const cache = await caches.open('wllama-local-models');\n    await cache.delete(url);\n  }\n}","import fs from 'fs';\nimport path from 'path';\n\nexport function copyWasmToPublic(publicDir: string): void {\n  const src = path.resolve(\n    require.resolve('@wllama/wllama/esm/index.js'),\n    '../../wasm/wllama.wasm'\n  );\n  const dest = path.join(publicDir, 'wllama', 'wllama.wasm');\n  \n  fs.mkdirSync(path.dirname(dest), { recursive: true });\n  fs.copyFileSync(src, dest);\n  console.log(`✓ Copied wllama.wasm to ${dest}`);\n}"],"mappings":";;;;;;;;AAAA,SAAS,cAAc;AAShB,IAAM,gBAAN,MAAoB;AAAA,EAKzB,YAAY,SAA8B,CAAC,GAAG;AAJ9C,SAAQ,SAAwB;AAChC,SAAQ,YAAoB;AAI1B,SAAK,SAAS;AAAA,MACZ,UAAU,OAAO,YAAY;AAAA,MAC7B,YAAY,OAAO,cAAc;AAAA,MACjC,MAAM,OAAO,QAAQ;AAAA,IACvB;AAAA,EACF;AAAA;AAAA,EAGA,mBAAqC;AACnC,QAAI,OAAO,gBAAgB,aAAa;AACtC,aAAO,EAAE,SAAS,OAAO,OAAO,gDAAgD;AAAA,IAClF;AACA,QAAI,EAAE,YAAY,SAAS;AACzB,aAAO,EAAE,SAAS,OAAO,OAAO,8CAA8C;AAAA,IAChF;AACA,UAAM,YAAY,CAAC,CAAE,UAAkB;AACvC,UAAM,WAAW,OAAO;AACxB,WAAO;AAAA,MACL,SAAS;AAAA,MACT;AAAA,MACA,qBAAqB;AAAA,IACvB;AAAA,EACF;AAAA;AAAA,EAGA,MAAM,UACJ,MACA,YAC0B;AAC1B,UAAM,MAAM,KAAK,iBAAiB;AAClC,QAAI,CAAC,IAAI,QAAS,QAAO,EAAE,SAAS,OAAO,OAAO,IAAI,MAAM;AAE5D,UAAM,KAAK,OAAO;AAElB,QAAI;AACF,mBAAa,EAAE;AAEf,YAAM,cAAc,EAAE,SAAS,KAAK,OAAO,SAAS;AACpD,WAAK,SAAS,IAAI,OAAO,WAAkB;AAC3C,mBAAa,EAAE;AAEf,YAAM,KAAK,aAAa,IAAI;AAC5B,mBAAa,EAAE;AAEf,YAAM,YAAY,CAAC,CAAE,UAAkB;AACvC,YAAO,KAAK,OAAe,UAAU,CAAC,IAAI,GAAG;AAAA,QAC3C,OAAO,KAAK,OAAO;AAAA,QACnB,GAAI,aAAa,KAAK,OAAO,aAAa,IACtC,EAAE,cAAc,KAAK,OAAO,WAAW,IACvC,CAAC;AAAA,MACP,CAAC;AAED,WAAK,YAAY,KAAK;AACtB,mBAAa,GAAG;AAEhB,aAAO,EAAE,SAAS,MAAM,YAAY,UAAU;AAAA,IAChD,SAAS,GAAQ;AACf,YAAM,KAAK,OAAO;AAClB,aAAO,EAAE,SAAS,OAAO,OAAO,EAAE,QAAQ;AAAA,IAC5C;AAAA,EACF;AAAA;AAAA,EAGA,MAAM,SAAS,KAA+C;AAC5D,QAAI,CAAC,KAAK,QAAQ;AAChB,aAAO,EAAE,SAAS,OAAO,OAAO,2CAA2C;AAAA,IAC7E;AAEA,QAAI;AACF,YAAM,KAAK,YAAY,IAAI;AAE3B,YAAM,WAAgD,CAAC;AACvD,UAAI,IAAI,OAAQ,UAAS,KAAK,EAAE,MAAM,UAAU,SAAS,IAAI,OAAO,CAAC;AACrE,eAAS,KAAK,EAAE,MAAM,QAAQ,SAAS,IAAI,OAAO,CAAC;AAEnD,YAAM,WAAW,MAAO,KAAK,OAAe,qBAAqB;AAAA,QAC/D;AAAA,QACA,YAAY,IAAI,aAAa;AAAA,QAC7B,aAAa,IAAI,eAAe;AAAA,QAChC,OAAO,IAAI,QAAQ;AAAA,QACnB,OAAO,IAAI,QAAQ;AAAA,MACrB,CAAC;AAED,YAAM,OAAe,UAAU,UAAU,CAAC,GAAG,SAAS,WAAW;AACjE,aAAO;AAAA,QACL,SAAS;AAAA,QACT;AAAA,QACA,QAAQ,KAAK,MAAM,YAAY,IAAI,IAAI,EAAE;AAAA,MAC3C;AAAA,IACF,SAAS,GAAQ;AACf,aAAO,EAAE,SAAS,OAAO,OAAO,EAAE,WAAW,oBAAoB;AAAA,IACnE;AAAA,EACF;AAAA;AAAA,EAGA,MAAM,SAAwB;AAC5B,QAAI,KAAK,QAAQ;AACf,UAAI;AAAE,cAAO,KAAK,OAAe,KAAK;AAAA,MAAG,SAAS,GAAG;AAAA,MAAC;AACtD,WAAK,SAAS;AAAA,IAChB;AACA,QAAI,KAAK,WAAW;AAClB,UAAI;AAAE,cAAM,KAAK,gBAAgB,KAAK,SAAS;AAAA,MAAG,SAAS,GAAG;AAAA,MAAC;AAC/D,WAAK,YAAY;AAAA,IACnB;AAAA,EACF;AAAA;AAAA,EAGA,IAAI,WAAoB;AACtB,WAAO,KAAK,WAAW;AAAA,EACzB;AAAA;AAAA,EAGA,IAAI,eAAuB;AACzB,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,MAAc,aAAa,MAA2B;AACpD,UAAM,MAAM,GAAG,OAAO,SAAS,MAAM,wBAAwB,KAAK,IAAI;AACtE,UAAM,QAAQ,MAAM,OAAO,KAAK,qBAAqB;AACrD,UAAM,MAAM,IAAI,KAAK,IAAI,SAAS,MAAM;AAAA,MACtC,SAAS,EAAE,gBAAgB,2BAA2B;AAAA,IACxD,CAAC,CAAC;AAAA,EACJ;AAAA,EAEA,MAAc,gBAAgB,UAAiC;AAC7D,UAAM,MAAM,GAAG,OAAO,SAAS,MAAM,wBAAwB,QAAQ;AACrE,UAAM,QAAQ,MAAM,OAAO,KAAK,qBAAqB;AACrD,UAAM,MAAM,OAAO,GAAG;AAAA,EACxB;AACF;;;AChJA,OAAO,QAAQ;AACf,OAAO,UAAU;AAEV,SAAS,iBAAiB,WAAyB;AACxD,QAAM,MAAM,KAAK;AAAA,IACf,UAAQ,QAAQ,6BAA6B;AAAA,IAC7C;AAAA,EACF;AACA,QAAM,OAAO,KAAK,KAAK,WAAW,UAAU,aAAa;AAEzD,KAAG,UAAU,KAAK,QAAQ,IAAI,GAAG,EAAE,WAAW,KAAK,CAAC;AACpD,KAAG,aAAa,KAAK,IAAI;AACzB,UAAQ,IAAI,gCAA2B,IAAI,EAAE;AAC/C;","names":[]}

package/package.json

@@ -0,0 +1,39 @@
+{
+  "name": "wllama-service",
+  "version": "1.0.0",
+  "description": "Framework-agnostic browser LLM service wrapper with WebGPU and WebAssembly support",
+  "main": "./dist/index.js",
+  "module": "./dist/index.mjs",
+  "types": "./dist/index.d.ts",
+  "exports": {
+    ".": {
+      "types": "./dist/index.d.ts",
+      "import": "./dist/index.mjs",
+      "require": "./dist/index.js"
+    }
+  },
+  "files": [
+    "dist",
+    "scripts",
+    "README.md"
+  ],
+  "scripts": {
+    "build": "tsup",
+    "dev": "tsup --watch",
+    "prepublishOnly": "npm run build",
+    "postinstall": "node scripts/postinstall.js"
+  },
+  "bin": {
+    "wllama-copy-wasm": "./scripts/postinstall.js"
+  },
+  "keywords": ["wllama", "llm", "webgpu", "webassembly", "llama", "browser-ai", "gguf"],
+  "license": "MIT",
+  "peerDependencies": {
+    "@wllama/wllama": ">=3.0.0"
+  },
+  "devDependencies": {
+    "@wllama/wllama": "^3.0.0",
+    "tsup": "^8.0.0",
+    "typescript": "^5.0.0"
+  }
+}

package/scripts/postinstall.js

@@ -0,0 +1,31 @@
+#!/usr/bin/env node
+const fs = require('fs');
+const path = require('path');
+
+// Find the consuming app's public directory
+const possiblePublicDirs = [
+  path.resolve(process.cwd(), '../../public'),      // monorepo
+  path.resolve(process.cwd(), '../../../public'),   // deeper monorepo
+  path.resolve(process.cwd(), 'public'),            // root install
+];
+
+const wasmSrc = path.resolve(
+  __dirname,
+  '../node_modules/@wllama/wllama/esm/wasm/wllama.wasm'
+);
+
+if (!fs.existsSync(wasmSrc)) {
+  console.warn('[wllama-service] Could not find wllama.wasm — run copy manually');
+  process.exit(0);
+}
+
+const publicDir = possiblePublicDirs.find(fs.existsSync);
+if (!publicDir) {
+  console.warn('[wllama-service] Could not find public/ directory — copy wllama.wasm manually');
+  process.exit(0);
+}
+
+const dest = path.join(publicDir, 'wllama', 'wllama.wasm');
+fs.mkdirSync(path.dirname(dest), { recursive: true });
+fs.copyFileSync(wasmSrc, dest);
+console.log(`[wllama-service] ✓ Copied wllama.wasm to ${dest}`);